In agriculture, soil fertility has a significant impact on farmer welfare. Soil fertility is the primary factor that determines crop production. Climate change, water resources, soil nutrients, and other environmental elements affect the soil fertility. In real time, the process of evaluating soil fertility is challenging and time-consuming. Many studies utilized Machine Learning-based classification and regression processes to determine soil fertility levels. Existing research has shown poor performance, time-consuming processes, and a lack of sustainable crop recommendations, leading to inefficient soil fertility classification and crop recommendations. To overcome these limitations, a novel Explainable Dual Attention-based Dense Capsule Dilated Depthwise Convolutional (EDA_DCDDC) method is introduced in this research. In addition, a sustainable agriculture recommendation system, based on fuzzy logic, is employed to regain soil fertility. Pre-processing methods such as Min-Max Normalization and the Improved Synthetic Minority Oversampling Technique (ISMOTE) are used to maintain data standards and balance minority classes, respectively. This model utilizes Dense Capsules to extract complex features. Then, spatial and channel feature modules are selected and incorporated to reduce dimensionality issues. Dilated Depthwise Convolution (DDC) is used to classify features into “Less Fertile”, “Fertile”, and “Highly Fertile” soils. After the classification process, fuzzy rules are employed to recommend season-based crops according to the soil fertility classes. Finally, SHapley Additive exPlanations (SHAP) are employed to reveal the results. In the experimental analysis, a well-defined soil fertility dataset is processed, producing a classification accuracy of 98.45% and an F1-score of 97.73%, resulting in more reliable performance.
Although convolutional neural networks report near-perfect accuracy in public finger vein datasets, their operational trustworthiness remains a critical, unevaluated blind spot. This study challenges the conventional reliance on accuracy as the primary metric for model viability. It is hypothesized that state-of-the-art models often achieve high scores not by learning robust vascular patterns, but by overfitting to spurious, non-biometric cues in the dataset background. To expose this vulnerability, an explainability-driven audit is conducted using Local Interpretable Model-Agnostic Explanations (LIME) on eight deep learning models across two benchmark datasets (SDUMLA and UTFVP). The results reveal a significant accuracy-reliability paradox. The highest-accuracy model, Densely Connected Convolutional Networks (DenseNet121-99.84%), is found to be functionally insecure, heavily relying on background artifacts for prediction. Conversely, Extreme Inception (Xception), despite lower raw accuracy, emerges as the most biometrically faithful architecture (Explainability Score: 89%), consistently targeting valid finger vein patterns. Unlike traditional performance-oriented studies, the primary novelty of this research lies in its explainability-driven audit approach, which systematically evaluates the trustworthiness of deep learning models in finger vein recognition, bridging the gap between high accuracy and model interpretability. This discovery demonstrates that high accuracy can mask significant security flaws, making models vulnerable to presentation attacks. Consequently, explainability is established not as a secondary feature, but as a mandatory security metric. It is proposed that future biometric evaluations must adopt interpretability analysis to ensure models characterize the biological trait rather than dataset noise.
This study addresses two critical challenges in social network–based consensus decision making (SNCDM): dynamic trust network analysis and consensus development design. A novel dynamic trust network analysis method is proposed to integrate direct and indirect trust relationships by adjusting direct trust levels based on opinion similarities through trust compensation mechanisms and estimating indirect trust levels via position-weighted trust propagation and path aggregation. A spectral clustering method is developed to partition decision makers into manageable subgroups for enhancing consensus development. This leads to the proposition of a SNCDM framework, which effectively addresses trust decay, mitigates over-trust effect, and preserves trust information integrity. A case study on prioritising barriers to the renewable energy transition in China is conducted to demonstrate the effectiveness of the proposed SNCDM framework. A comparative analysis shows that the proposed framework performs better in trust information preservation and consensus acceleration than comparable approaches. Overall, this study offers a practical solution for SNCDM in which trust relationships embedded in social networks play a central role in consensus development.
Online travel platforms such as Trip.com and Agoda have made it easier to modify or cancel hotel reservations, increasing uncertainty for hotel revenue management. Cancellations can increase room vacancy, destabilize revenue, and raise operating costs. To address these challenges, this study proposes a feature-interaction-enhancement framework that enriches reservation data with interpretable cross-dimensional features designed to capture nonlinear relationships between booking behavior and cancellation risk. Unlike hybrid prediction models that jointly optimize a Bayesian network (BN) and a classifier, the proposed framework uses the BN solely as an auxiliary feature extractor. It generates two meta-features — the posterior cancellation probability and its entropy — which are concatenated with the original and interaction features for downstream prediction. We systematically compare several optimized machine-learning models, including Decision Tree, Random Forest, Gradient Boosting, and XGBoost, and apply SHapley Additive exPlanations (SHAP) to the best-performing model to quantify feature contributions and derive operational insights. The experimental results show that the feature-enhanced framework outperforms baselines that rely solely on the original predictors. The resulting risk estimates can help hotels identify high-risk bookings and design targeted interventions, provided they are appropriately calibrated and operationally validated.
Advancements in natural language processing (NLP) and sentiment analysis (SA) have revolutionized humancomputer interactions and opened up numerous opportunities across various industries. However, it is important to note that no sentiment analysis lexicon is without flaws or universally applicable. Each lexicon is influenced by its creation process, leading to variations in coverage and inherent biases. This emphasizes the need for careful selection of appropriate lexicons in the research and development of lexicons, as the diversity of available lexicons can significantly impact the success of sentiment models. This requires the consideration of several factors, such as the characteristics of the available datasets, the lexicons used for evaluation, the evaluation criteria drawn from the literature, and their relevance to the task. Thus, evaluating and benchmarking SA lexicons often poses a complex multicriteria decision-making (MCDM) problem because of these issues. To our knowledge, a practical MCDM approach for systematically addressing this MCDM problem remains underdeveloped in the SA literature. To bridge this gap, this study introduces an integrated framework based on the utilization of the fuzzy-weighted zero-inconsistency (FWZIC) method for weighting the evaluation criteria and the fuzzy decision by opinion score method (FDOSM) for prioritizing lexicon alternatives. The proposed framework is applied to 28 ASA lexicons as a case study. The evaluation framework focuses on four main criteria and 28 subcriteria. The weighting results show that the labelling type category acquired most of the main weights (0.269762809) compared with the labelling target (0.230973301). For the sub-LW weights, the (manual labelling) subcriteria acquired the most weight (0.364022932) compared with (No. of Authors), which acquired the least weight (0.07793705). Among the 28 lexicon alternatives, TEAD achieved the highest evaluation score, whereas ‘TSAC’ ranked lowest on the basis of the weighting criteria. A sensitivity analysis was employed to evaluate the proposed approach. For the implications of this research, the proposed framework offers valuable insights for researchers and practitioners in selecting optimal lexicons for Arabic sentiment analysis. Furthermore, this framework has the potential to be applied to other NLP applications beyond sentiment analysis.
